CN112949186B - Method for predicting wax precipitation point temperature of wax-containing crude oil based on SSA-LSSVM model - Google Patents

Abstract

The invention discloses a method for predicting wax precipitation point temperature of wax-containing crude oil based on an SSA-LSSVM model, which optimizes a penalty parameter gamma and a kernel function width sigma in a Least Square Support Vector Machine (LSSVM) model by utilizing a Sparrow Search Algorithm (SSA) ² Establishing a wax precipitation point temperature SSA-LSSVM prediction method for the wax-containing crude oil; the method utilizes the SSA-LSSVM model to predict the wax precipitation point temperature of the waxy crude oil, has high prediction precision and strong generalization capability, is not easy to fall into the problem of local optimum, can dynamically provide reference data for engineering, and provides certain help for the reasonable development of the waxy oil and gas reservoir.

Description

Method for predicting wax precipitation point temperature of wax-containing crude oil based on SSA-LSSVM model

Technical Field

The invention relates to the technical field of petroleum and natural gas exploration and development, in particular to a method for predicting wax precipitation point temperature of wax-containing crude oil based on an SSA-LSSVM model.

Background

In the process of oil and gas reservoir exploitation, along with the change of conditions such as temperature, pressure, fluid flow speed and the like, solid-phase substances mainly containing paraffin in oil and gas reservoir fluid can be gradually separated out and deposited, the paraffin deposited in a reservoir layer can reduce the porosity and permeability of the reservoir layer, the paraffin deposited on the wall surfaces of a shaft and ground equipment can reduce the effective flow area and even completely block the shaft and the ground flow pipeline, the productivity of an oil and gas well is reduced, and even production is stopped in severe cases, so the influence and the harm brought by paraffin deposition on normal oil production cannot be ignored.

The paraffin wax existing in crude oil or condensate gas system has very fine coarse-crystal wax and micro-crystal wax, when the temperature is lower than the wax precipitation point temperature, wax crystal will be deposited, and if the shearing stress acting on the paraffin wax is lower than the shearing strength of the deposited paraffin wax, the flowing fluid can be blocked, and the blockage is caused. Paraffin deposition occurring in oil and gas reservoirs is always a major technical point difficult point to solve, is a necessary research subject needing to be developed in the petroleum industry at present, and needs to be researched by combining experiments and fluid thermodynamic theory due to higher difficulty. The deep research on the paraffin deposition problem not only can solve the problem of solid phase deposition in a shaft, a gathering and transportation pipeline and a stratum, which puzzles people for a long time in the process of exploiting an oil and gas field, but also can provide scientific research means and scientific prediction methods for the efficient development of improving the recovery ratio of the oil and gas reservoir by high-condensation and high-viscosity oil reservoir, a deep gas reservoir, a condensate gas reservoir and gas injection. In recent years, artificial intelligence means is developed vigorously, and the intelligent algorithm model can well predict the wax precipitation point temperature of the wax-bearing crude oil without analyzing a complex deposition process. However, the existing intelligent algorithm model has the defects that the prediction precision is not enough, and the local optimization is easy to fall into. Therefore, a new model method is needed to predict the wax precipitation point temperature of the waxy crude oil, so as to lay a foundation for reasonable development of the waxy oil and gas reservoir.

Disclosure of Invention

In order to solve the technical problems, the invention aims to provide a method for predicting the wax precipitation point temperature of the waxy crude oil based on an SSA-LSSVM model. The model has the advantages of high prediction precision, strong generalization capability and difficulty in falling into the problem of local optimum, and can provide effective reference data for experimental analysis and actual engineering.

In order to achieve the purpose of the invention, the invention adopts the technical scheme that:

the method for predicting the wax precipitation point temperature of the wax-containing crude oil based on the SSA-LSSVM model comprises the following specific steps:

step 1: determining input variables and output variables for modeling by using a Pearson correlation method; dividing experimental data into two parts, and carrying out training data and prediction data according to the proportion of 7;

step 2: establishing a prediction model of wax precipitation point temperature of the wax-containing crude oil by using an LSSVM algorithm;

and step 3: normalizing the original data, and utilizing data of optimized parameters of a Sparrow Search Algorithm (SSA) and sample data of training and testing of a Least Square Support Vector Machine (LSSVM) prediction model;

and 4, step 4: initializing and setting parameters of an SSA algorithm and an LSSVM prediction model: binary coding is carried out by utilizing the collected wax precipitation point temperature data to generate an initial generation population, namely an initial LSSVM model, then the model is trained, LSSVM prediction model parameters are obtained through SSA algorithm optimization, a sparrow search algorithm least square support vector machine (SSA-LSSVM) prediction model is established, and wax precipitation point temperature prediction is carried out on a test sample;

and 5: and (4) analyzing and comparing the prediction result obtained in the step (4) with the test value by adopting theories of abnormal point detection and the like, and verifying the correctness of the prediction value and the application range of the model.

Compared with the prior art, the technology adopted by the invention has the following advantages:

according to the invention, the SSA-LSSVM wax-bearing crude oil wax precipitation point temperature prediction model established by the LSSVM is adopted, so that the defects of a common model are overcome, and the model has the advantages of strong generalization capability, high prediction precision, good adaptability and difficulty in falling into the problem of local optimization; the invention optimizes LSSVM model parameters by selecting an SSA algorithm, and has the advantages of processing the problem of discrete variables, processing continuous variables and nonlinear targets, not needing gradient information to constrain functions and the like.

Drawings

FIG. 1 is a basic schematic diagram of a support vector machine

FIG. 2 is a basic method flow diagram of the present invention

FIG. 3 is a graph showing the results of using SSA-LSSVM model to predict the wax point temperature of waxy crude oil

FIG. 4 is a graph showing the comparison result of the average AARD between the prediction models

FIG. 5 is a diagram showing the results of abnormal point detection in the SSA-LSSVM model

Detailed Description

The technical scheme of the invention is described in detail in the following with reference to the attached drawings:

background of the invention

A Support Vector Machine (SVM) is a generalized linear classifier for binary classification of data in a supervised learning mode, a decision boundary of the SVM is a maximum edge distance hyperplane for solving a learning sample, nonlinear classification can be performed through a kernel method, and the SVM is one of common kernel learning methods. Compared with other prediction methods, the SVM has the advantages that the structure risk is small, and the SVM is suitable for data processing with small samples and high dimensionality. Prediction accuracy of SVM depends on the pineRelaxation variable ζ _i And a normalization constant C, a normalization constant C and a relaxation variable ζ _i The prediction accuracy of the sample is directly influenced by the change of the input data. The direction of various algorithm improvements becomes how to affect C and ζ _i The value of (a). In order to make SVMs computationally convenient, least Squares Support Vector Machines (LSSVMs) have come into play. The LSSVM is developed by applying a least square correlation principle on the basis of the SVM, and has the following two characteristics: firstly, an inequality constraint in the SVM is changed into an equality constraint by adopting a quadratic programming method, and a quadratic programming problem is replaced by solving a linear equation set problem; and secondly, the defects of rough data set, insufficient data quantity and the like which are difficult to solve by a conventional prediction method are overcome, and the method is particularly suitable for small sample prediction. Obviously, the LSSVM has the advantages that the operation process of the LSSVM is simplified, and the operation efficiency is obviously improved.

Based on the principle of minimizing structural risk and the basis of statistical theory, the SVM is suitable for the research of linear classification problems, and can also be applied to the regression calculation problem of nonlinear data and the fitting aspect of unknown function curves by solving a convex quadratic optimization problem. FIG. 1 is a basic block diagram of a support vector machine. In form, the SVM has an input end and an output end of data information as well as a feedforward neural network; however, the neurons of the neural network are replaced by the kernel function of the SVM, so that the two principles of operation are different. The kernel function in the SVM is used for converting a low-dimensional input vector into a higher-dimensional vector space, and after the kernel function is converted, the SVM can select a proper optimization algorithm to execute linear regression or classification calculation.

The method for predicting the wax precipitation point temperature of the wax-containing crude oil based on the SSA-LSSVM model comprises the following specific steps:

the first step is as follows: in order to ensure the accuracy and the effectiveness of the model, 4 groups of wax-containing crude oil wax precipitation point temperature data (107 data points) published at home and abroad are selected as experimental data of the method, and the data are normalized by a mapminmax function in order to improve the prediction precision of the model due to inconsistent order of magnitude of the experimental data. 70% of each of the 4 groups of data was used as training samples (Train) and 30% of the remaining data was used as prediction samples (Test); determining the main influencing factors of the wax precipitation point temperature of the wax-containing crude oil by utilizing a Pearson correlation method, namely P and MW; therefore, the invention selects the variables as input variables and the wax precipitation point temperature as output variables; the mapping used by the mapminmax function is:

the second step is that: establishing a prediction model of wax precipitation point temperature of the wax-containing crude oil by using an LSSVM algorithm; the specific process is as follows:

firstly, an input variable is recorded as x, an output variable is recorded as y, and the wax precipitation point temperature data are divided into two parts: training data and predictive data. Assume a sample number set { (x) _i ,y _i ) I =1,2,3, \8230, N is a column vector of N dimensions, where N is the number of samples, x _i Is an input value of the ith data, y _i For the output numerical value of the ith data, introducing a kernel function into the SVM algorithm, and mapping the input sample space with nonlinear change into a high-dimensional feature space, thereby solving the problem that the sample data can be linearly divided. After kernel transformation, the decision function is of the form:

in the formula, ω is a weight coefficient;

b-offset, b ∈ R.

By finding the optimal ω and b, the confidence range is minimized under the condition that the Laplace loss function is fixed and unchanged, and the optimization problem is converted into:

in the formula, x _i -the ith learning sample input value;

y _i -the ith learning sample output value;

thus the problem is converted into an evaluation1/2||ω|| ² The training error is the constraint condition of the optimization problem, so that the corresponding regression function is obtained through changing, and the optimization method has good generalization performance. If the constraint condition is not satisfied under the condition that a certain error is allowed, adding a relaxation variable zeta _i 、ζ _i ^* The optimization problem then turns into:

in the formula, ζ _i 、ζ _i ^* -a relaxation variable;

c is a normalization constant, and C is more than or equal to 0.

To more conveniently solve the equation, we change the previous convex quadratic programming problem, thereby introducing the lagrangian function:

in the formula, alpha _i 、α _i *、η _i 、η _i * -Lagrange multiplier pairs for each sample.

From the above equation, the corresponding solving problem is as follows:

according to the Karush-Kuhn-Tucker condition, at the optimal point, the lagrange multiplier is multiplied by the corresponding constraint to be zero, that is:

can derive a _i ×α _i *＝0，α _i And alpha _i * Must have a value of zero, and the offset b can be calculated by the standard support vector machine, i.e.:

in order to further improve the calculation accuracy, it is a common practice to calculate the corresponding offset of the standard SVM, and then calculate the average value, then:

where Nnsv is the number of standard support vectors.

Therefore, from equations (8) and (9), the final regression function expression is:

LSSVM and SVM have the same structure, and have an input layer and an output layer, wherein a hidden layer comprises a kernel for converting low-dimensional input data into high-dimensional input data, and the input vector is converted into a false high-dimensional vector of a feature space through the kernel, so that the high-dimensional space vector differentiability and the low-dimensional space computability are realized, but the working principles of the LSSVM and the SVM are different, the SVM is based on inequality constraint, and the LSSVM is based on equality constraint, so that the original problem is converted from a quadratic programming problem into a linear equation problem of a linear KKT system.

The basic idea of LSSVM regression is to select a training set (x) by transforming the nonlinear regression problem of the low-dimensional space into the linear regression problem of the high-dimensional feature space _i ,y _i ) I =1,2, \8230;, l, the regression function of LSSVM is:

according to the principle of minimizing structural risk, the optimization goal of LSSVM can be expressed as:

introducing Lagrange multiplier alpha _i Then the Lagrange polynomial of the dual problem of equation (12) is:

substituting formula (13) into the Karush-Kuhn-Tucker (KKT) condition yields:

thereby replacing the solved solution optimization problem with the linear system of equations problem in solving equation (14).

In the formula, I = [1,2, \8230;, l] ^T ，α＝[α ₁ ,α ₂ ,…,α _l ] ^T ，y＝[y ₁ ,y ₂ ,…,y _l ] ^T ，A＝ZZ ^T +γ ^-1 I，

i＝1,2,…,l。

Finally, the LSSVM regression model becomes:

in the formula, K (x) _i And x) is a kernel function of the LSSVM, and the kernel function of the least squares support vector machine is as follows: polynomial function, RBF kernel function, sigmoid function. A large number of researches show that the RBF kernel function is better adopted in the regression prediction, and the expression is as follows.

K(x _i ,x)＝exp(-‖x-x _i ‖ ² /2σ ² ) (17)

In the formula, σ ² The kernel function width reflects the radius encompassed by the boundary closure. In the regression model of LSSVM, the penalty parameter γ and the kernel function width σ ² Are the two parameters that affect the LSSVM performance the most. The invention adopts a sparrow search algorithm to optimize gamma and sigma ² Two parameters.

The third step: normalizing the original data, and optimizing the data of parameters by utilizing a Sparrow Search Algorithm (SSA) and sample data of training and testing a prediction model of a least square support vector machine; in the present invention, the penalty parameter γ and the kernel function width σ are required ² Optimized to obtain optimal gamma and sigma ² 173.345 and 0.7218 respectively; the specific optimization process is shown in fig. 2.

The fourth step: initializing and setting parameters of a prediction model of a Sparrow Search Algorithm (SSA) and a Least Square Support Vector Machine (LSSVM): binary coding is carried out by utilizing the collected wax precipitation point temperature data to generate an initial generation population, namely an initial Least Square Support Vector Machine (LSSVM) model, then the model is trained, least Square Support Vector Machine (LSSVM) prediction model parameters are obtained through optimization of a Sparrow Search Algorithm (SSA), a sparrow search algorithm least square support vector machine (SSA-LSSVM) prediction model is established, and wax precipitation point temperature prediction is carried out on a test sample; the results of SSA-LSSVM model to predict the wax point temperature of waxy crude are shown in FIG. 3. As can be seen from FIG. 3, the predicted values of the SSA-LSSVM model are mostly uniformly distributed around the 45-degree line, and the model accuracy is high.

The fifth step: and (4) analyzing and comparing the prediction result obtained in the step (4) with the test value by adopting theories of abnormal point detection and the like, and verifying the correctness of the prediction value and the application range of the model.

The absolute average error (AARD) of experimental and calculated wax-out point temperatures is defined as follows:

in order to facilitate the correlation calculation of the AARD result of each group of wax precipitation point temperature of the analysis model, the AARD formula is used for calculation after the predicted wax precipitation point temperature is obtained, and the average error of the predicted wax precipitation point temperature of the SSA-LSSVM model is 4.73% and the maximum average error is 24.15% under the conditions that the temperature is 270-320K and the pressure is 0-70 MPa.

In order to verify the effectiveness of the models, an LSSVM is used for comparison test, the test data is unchanged, and the comparison result of the absolute average error AARD corresponding to each prediction model is shown in fig. 4. Fig. 4 is a graph showing the comparison result of the average AARD corresponding to each prediction model.

And calculating a hat value H and a standard deviation SR of the SSA-LSSVM model by using an abnormal point detection theory to detect abnormal points, wherein all data points are distributed in a region controlled by H not less than 0 and H not more than 3 and R not less than 3, and no abnormal point exists, so that the model is effective, and the reliability and the accuracy of the model are further explained.

The SSA algorithm is used for carrying out parameter optimization on the LSSVM model, a wax precipitation point temperature prediction model based on the SSA-LSSVM is established, sensitivity analysis calculation shows that the prediction result of the SSA-LSSVM model is superior to that of the LSSVM model and an improved Hsu model, and anomaly point detection is carried out, so that the effectiveness and the accuracy of the method are fully demonstrated, and accurate prediction of data can be excellently realized.

Claims (1)

1. A method for predicting wax precipitation point temperature of wax-containing crude oil based on an SSA-LSSVM model is characterized by comprising the following specific steps:

step 1: determining input variables and output variables for modeling by using a Pearson correlation method; dividing experimental data into two parts, and performing training data and prediction data according to the proportion of 7: 3;

step 2: establishing a prediction model of the wax precipitation point temperature of the wax-bearing crude oil by using an LSSVM algorithm;

K(x _i ,x)＝exp(-||x-x _i || ² /2σ ² )

in the formula: k (x) _i And x) is a kernel function of the LSSVM; sigma ² Is the kernel function width; b is an offset; x is the number of _i Inputting a value for the ith learning sample; y is output data;

and step 3: carrying out normalization processing on the original data, and optimizing parameter data and LSSVM prediction model training and testing sample data by utilizing an SSA algorithm;

and 4, step 4: initializing and setting parameters of an SSA algorithm and an LSSVM prediction model: binary coding is carried out by utilizing the collected wax precipitation point temperature data to generate an initial generation population, namely an initial LSSVM model, then the model is trained, LSSVM prediction model parameters are obtained through SSA algorithm optimization, a sparrow search algorithm least square support vector machine (SSA-LSSVM) prediction model is established, and wax precipitation point temperature prediction is carried out on a test sample;

and 5: and (4) analyzing and comparing the prediction result obtained in the step (4) with the test value by adopting an abnormal point detection theory, and verifying the correctness of the prediction value and the application range of the model.

Images